Stewart Vickers AI | Jan 0.8.4 is INSANE (free, offline, business-safe)

Six point seven million people already stopped paying for ChatGPT and switched to Jan, a free open-source desktop app that runs Llama, Qwen, Mistral, DeepSeek and Gemma on your own laptop with nothing leaving the machine.

Show Notes

Watch the video: https://www.youtube.com/watch?v=hj3VVEuXlyc

What you get in this one:
- Why Jan 0.8.4 (published July 23) is the release that made it compliance-safe: API keys moved out of browser storage into the OS keyring
- The exact model to pick for 16GB, 32GB and 64GB laptops, and the workflows I already moved off ChatGPT Team onto Jan
- How to keep GPT-5 and Claude for hard jobs while sensitive prompts stay on the local model, from the same interface
- The OpenAI-compatible localhost endpoint that lets your existing scripts point at Jan by changing one URL
- MCP tool support with an app-enforced approval popup, and the handlebars RCE patch that shipped in the same release

Jan download: https://jan.ai
Jan on GitHub (44,645 stars, 3,000 forks): https://github.com/janhq/jan
0.8.4 release notes: https://github.com/janhq/jan/releases

Chapters below. Companion configs, the model I recommend for a 16GB MacBook and the operator setup I use inside Jan sit in the pinned comment.

Chapters:
00:00 6.7M downloads: what people switched to
00:12 Jan: 57MB, free, offline
00:33 How Jan works: model on the laptop
01:15 Seats Jan retires: ChatGPT Team, Claude Team, Enterprise
01:55 The headline answer for business owners
02:18 Free Skool + pinned comment
02:48 GitHub numbers on Jan 0.8.4
03:10 Jan 0.8.4: keys moved to the OS keyring
03:47 Three things Jan does (local + hosted + localhost API)
04:16 The two-lane operator setup
04:44 MCP tool approval + handlebars RCE patch
05:14 Local model quality vs GPT-5, honestly
05:30 RAM to model-size guide: 16GB, 32GB, 64GB
05:58 3 cents an hour vs $30 a seat
06:15 Workflows I already moved off ChatGPT
06:38 What Jan does NOT replace
07:03 Verify it yourself: open jan.ai
07:26 Install it this afternoon

Audio edition of the Stewart Vickers AI YouTube video of the same name. Full transcript on this episode page.

What is Stewart Vickers AI?

The audio edition of the Stewart Vickers AI YouTube channel. Each episode takes one AI release - a model, a tool, a price cut - and translates it into money for business owners: what got cheaper, who got undercut, and what to do about it this week. Operator credibility over theory, specific numbers over vibes. Every episode links the original video and the free Church of SEO Jesus community where the prompts and workflows live.

7 million people already stopped paying for ChatGPT. Most business owners have never heard of the thing they switched to. It is called JAN. 57 megabytes. Free. Open source.

Every prompt you type stays on your laptop. No cloud. No account. No monthly bill. 8. 4 shipped a change in July that makes it business safe for the first time.

I will show you what the change is, which laptop it runs on, and which paid tools it retires today. JAN is a free desktop app from a group called Menlo Research. You install it. You pick a Modal from a list. You type. The Modal runs on your own CPU or GPU.

Nothing you type ever leaves the machine unless you explicitly tell it to. One sentence and the whole product fits in it. 7 million, while most of your competitors are still pasting client contracts into a browser tab that logs every keystroke to a server they will never audit. Here is what JAN can retire. ChatGPT Team. $30 PR seat PR month.

Claude Team. $30 PR seat PR month. ChatGPT Enterprise. $60 a seat. Minimum contract. Any small business paying for 5 seats is at $1,800 a year for a chat window.

JAN runs the same class of open models those companies charge you to access indirectly. Lamar from Meta. Qwen from Alibaba. Mistral. DeepSeek. GemR from Google.

KeyMe from Moonshot AI. Any of them offline on your laptop for $0. Here is the headline answer. If your team uses AI to summarize emails, draft first-pass proposals, rewrite marketing copy, or read contracts and PDFs before you sign them, you can do all of it on JAN today at zero recurring cost, and your client data never leaves your building. Zero recurring cost. Client data stays local, and it works on a laptop you already own.

Quick ask. My channel covers every free local Modal route the day it ships, so subscribe, and the next one finds you instead of you finding it weeks late on an invoice. The download link. The Modal I recommend for a 16GB MacBook, and the operator configs I run inside JAN all sit in the pinned comment, and inside the church of SEOJesus, my free school community, where I answer when your local setup refuses to open a PDF at midnight. Free to join. Link in the description.

I pulled the numbers from GitHub directly because a counter on a homepage is easy to fake. 4. Published July 23rd. 105,000 Windows downloads on that release alone. 75 million across every release the project has ever cut. 44,658 GitHub stars.

3,000 forks. Not a weekend project. The change that made me record this video is one line in the release notes. 4, any API keys you gave JAN for hosted models sat in browser storage inside the app. 4, they moved to the operating system keyring, the same vault your Mac uses for Wi-Fi passwords. Boring change.

Big consequence. It means JAN graduated from an enthusiast toy to something a compliance officer can actually approve. I sent the installer to two founders yesterday, neither called back with a security question. Both installed it inside an hour. Three things JAN actually does. 1.

It runs open models on your machine, fully offline, nothing leaves. 2. It also plugs into ChatGPT, Claude and Gemini if you want to root hard jobs to a hosted Modal from the same interface, with the IPI key held in your operating system. 3. It exposes an OpenAI-compatible endpoint on localhost, so any script your developer already wrote against the OpenAI IPI points at JAN by changing one URL. Same code, private inference, zero migration.

You get the sensible operator setup for free. Sensitive prompts, contracts, staff issues, financial numbers, the offer letter you have not sent yet, go to the local Modal, long-form writing where quality matters more than privacy, goes to Claude or GPG-5, through JAN with the API key stored in your operating system rather than a browser tab, two lanes, one interface, the right prompt goes to the right Modal without you having to think about it every time. 4 release also added support for MCP servers, which is the standard Anthro pick and OpenAI now both use for connecting a Modal to your own tools. It means JAN can hit your calendar, your CRM, your file store, with a tool approval pop-up asking you every time, and the pop-up is enforced by the app, not the Modal, cannot be prompt injected out of. The same release patched a remote code execution vulnerability in a bundled library called Handlebars. Someone is auditing this code, it matters.

Now the honest part. Local Modal quality is not the same as GPG-5. A 7 billion parameter LAMR running on a MacBook is roughly as useful as chat GPG was in early 2024, still enough to eat most of the tasks your team does every day. On a MacBook with 16 gigabytes of unified memory, you can run 7 and 8 billion parameter models at reading speed. On Thirty-two gigabytes, you can run 12 to 14 billion parameter models comfortably. On a fresh MacBook Pro with 64 gigabytes, 30 billion parameter models.

An 8-year-old business laptop with 8 gigabytes of RAM will struggle with anything above a 3 billion parameter Modal, and I am not going to lie to you about that. Electricity. A MacBook Air pulling 20 to 30 watts under sustained inference costs you under 3 cents an hour to run at UK residential rates. Estimated, not measured on your bill. A hosted seat costs you $30 a month whether you use it once or a thousand times. Here are the workflows I already moved off chat GPG and onto Jan on my own machine.

Summarizing client onboarding notes, drafting first pass proposals, rewriting cold email copy, reading contracts and technical PDFs before I sign them, naming things, turning meeting notes into action items. All of it stays on this laptop now, and my monthly AI bill dropped by the price of a decent dinner. What I still pay for. Long-form writing where the Modal has to hold a whole 90-page document in its head. Deep code review across a repository. Anything I want the very best current Modal on the planet for, Jan does not replace that.

Jan retires the middle, which is where most of your recurring bill actually lives. Playbooks and configs sit in the church of SEO Jesus, link below. ai yourself right now. 7 million counter next to it. Look at the operating system icons underneath. Install it on a spare afternoon.

Try it on one task you would normally paste into chat GPG. See if you notice the difference. That is the proof, not me telling you. 7 million people got there before you. 4 is the release that made it business safe. Install it this afternoon.